Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Expanding Universal Machine Learning Interatomic Potentials to 97 Elements Towards Nuclear Applications

Naoya Kuroda, Kenji Ishihara, Tomoya Shiota, Wataru Mizukami

Abstract

Machine learning interatomic potentials (MLIPs) evaluate potential energy surfaces orders of magnitude faster while maintaining accuracy comparable to first-principles calculations, and universal MLIPs that cover most of the periodic table are becoming increasingly commonplace. However, existing large-scale datasets have limited or no coverage of heavy elements such as minor actinides crucial in the nuclear field, and universal MLIPs are typically limited to 89 elements. Here, we constructed a heavy element dataset HE26 containing minor actinides, based on experimental and computational literature data. By integrating this with existing molecular and crystal datasets, we developed an open-source universal MLIP covering 97 elements, the broadest elemental coverage to date. The resulting model showed strong performance on the inorganic MPtrj and organic OFF23 test sets and promising accuracy on HE26. The dataset and model open a pathway toward the development of energy resources and the design of novel materials, such as actinide-based high-entropy ceramics, in the nuclear field.

Expanding Universal Machine Learning Interatomic Potentials to 97 Elements Towards Nuclear Applications

Abstract

Machine learning interatomic potentials (MLIPs) evaluate potential energy surfaces orders of magnitude faster while maintaining accuracy comparable to first-principles calculations, and universal MLIPs that cover most of the periodic table are becoming increasingly commonplace. However, existing large-scale datasets have limited or no coverage of heavy elements such as minor actinides crucial in the nuclear field, and universal MLIPs are typically limited to 89 elements. Here, we constructed a heavy element dataset HE26 containing minor actinides, based on experimental and computational literature data. By integrating this with existing molecular and crystal datasets, we developed an open-source universal MLIP covering 97 elements, the broadest elemental coverage to date. The resulting model showed strong performance on the inorganic MPtrj and organic OFF23 test sets and promising accuracy on HE26. The dataset and model open a pathway toward the development of energy resources and the design of novel materials, such as actinide-based high-entropy ceramics, in the nuclear field.
Paper Structure (26 sections, 2 equations, 7 figures, 2 tables)

This paper contains 26 sections, 2 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Results
  3. Cross-domain Accuracy
  4. Detailed Evaluation on HE26 Dataset
  5. Elemental Solids and Binary Oxides
  6. Multi-Component Systems
  7. Compositionally Complex Fluorite Oxides
  8. Lattice Thermal Conductivity
  9. Discussion
  10. Conclusion
  11. Methods
  12. HE26 Dataset
  13. Basic Heavy Element Subset
  14. Complex Heavy Element Subset
  15. Compositionally Complex Fluorite Oxide Subset
  16. ...and 11 more sections

Figures (7)

  • Figure 1: Overview of the 97-element universal dataset and the MACE-Osaka26 model. (a) Periodic table showing the element distribution (log scale) of the combined dataset consisting of MPtrj, OFF23, and our HE26 dataset. Red boxes indicate the heavy elements newly added in this study. (b) Schematic of the training pipeline integrating molecular (OFF23) and crystal (MPtrj, HE26) data via the total energy alignment protocol.
  • Figure 2: Accuracy comparison between MACE-Osaka24 and MACE-Osaka26 on cross-domain systems. (a) Evaluation on the MPtrj training set representing crystalline systems, and (b) the OFF23 training set representing molecular systems, using both MACE-Osaka24 (orange) and MACE-Osaka26 (blue) models. (c) Results for the HE26 training set containing heavy elements. Note that only MACE-Osaka26 is shown in (c) as it incorporates eight heavy elements not supported by the MACE-Osaka24 model. Mean absolute errors (MAE) and root-mean-square errors (RMSE) are provided in eV/atom. Here, the main difference in the architecture between MACE-Osaka24 and MACE-Osaka26 is that their graph cutoff values are 4.5 Å and 6.0 Å, respectively.
  • Figure 3: Parity plots of DFT-calculated versus MLIP-predicted lattice parameters for heavy element systems. (a) The Basic Heavy Element (BHE) subset, comprising 42 unique structures. (b) The Complex Heavy Element (CHE) subset, comprising 133 unique structures. Three structures were excluded because their structural optimizations did not converge. The specific systems illustrated in the insets represent typical experimentally synthesized examples. The lattice parameter of each structure is defined throughout this paper as the average length of the unit cell lattice vectors.
  • Figure 4: Evaluation of MACE-Osaka26 on compositionally complex fluorite oxides. (a) Comparison of predicted lattice parameters with DFT reference values and experimental data across the actinide series (Th to Cf). The bottom panel shows the error of MACE-Osaka26 relative to DFT. (b) Predicted lattice parameters for fluorite-structured oxides containing various rare metals and lanthanides. (c) Parity plot of MLIP-predicted versus DFT-calculated lattice parameters, color-coded by the number of constituent elements (excluding oxygen). (d) Root-mean-square error (RMSE) and (e) error distribution relative to DFT as a function of the number of constituent elements.
  • Figure 5: Lattice thermal conductivity ($\kappa_L$) of actinide oxides predicted by MACE-Osaka26. (a) Temperature dependence of $\kappa_L$ for a series of actinide dioxides ($An$O_2, $An$ = Th, Pa, U, Np, Pu, Am, Cm, Cf). The solid lines with symbols represent the MLIP results calculated by solving the Wigner transport equation via phono3py, while the open symbols denote experimental data from the literature where available. (b) Comparison of $\kappa_L$ between the pure binary oxides and the ternary solid solution Am_0.25U_0.75O_2. The experimentally obtained lattice thermal conductivities for pure UO_2 fink2000uo2, AmO_2 bakker1998thermal, and the solid solution Am_0.2U_0.8O_2 chakraborty2025thermal are plotted as dotted curves to validate the MLIP predictions.
  • ...and 2 more figures